Wind turbine life prediction method based on empirical mode decomposition and hybrid neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-13
- Publication Date
- 2026-08-11
AI Technical Summary
然而，如何减小噪声干扰并获得较好的预测结果，仍然是风电机组寿命预测面临的核心挑战
本发明通过改进的经验模态分解（CEEMDAN）和混合神经网络（CNN-LSTM +Attention Mechanism）相结合的方法，有效解决了传统寿命预测方法在风电机组寿命预测中存在的抗噪能力不足、复杂工况适应性差等问题。该方法不仅显著提高了预测精度，还增强了模型的鲁棒性和可解释性，为风电行业的智能运维提供了有力支持。
Smart Images
Figure CN122548249A_ABST
Abstract
Description
Technical Field This invention belongs to the field of wind turbine remaining life prediction and health management technology, specifically involving a wind turbine life prediction method based on empirical mode decomposition and hybrid neural network. Background Technology
[0001] Against the backdrop of a global energy structure transition towards clean and low-carbon energy, the large-scale application and safe, stable operation of wind power equipment are directly related to the clean energy substitution process and the optimization and upgrading of the energy structure, serving as a crucial support for promoting high-quality development of the energy industry. Currently, with the continuous expansion of installed wind power capacity and the increasing service life of wind power equipment, the long-term reliable operation of core components faces severe challenges. Key components such as gearboxes and generators operate under complex and harsh conditions for extended periods, making them prone to wear, fatigue, and performance degradation. Traditional methods relying on reactive measures and regular maintenance are insufficient for early fault prediction and accurate trend forecasting, often leading to undetected hazards and unplanned outages. This not only significantly increases the total life-cycle cost but also severely impacts the power generation efficiency and operational safety of wind farms, hindering the full release of wind power equipment's effectiveness. Therefore, developing highly reliable and accurate wind turbine condition prediction and early fault warning technologies, shifting from "reactive maintenance" to "proactive prediction," has become crucial for overcoming industry bottlenecks, ensuring the long-term safe and efficient operation of wind power equipment, and promoting the high-quality development of the wind power industry.
[0002] In recent years, with the development of big data and artificial intelligence technologies, data-driven life prediction methods have gradually received widespread attention [Yao Liang, Guo Gang. Data-driven fault diagnosis method for wind turbine gearbox oil circuit system [J]. New Industrialization, 2019, 9 (04): 6-11.]. The wind turbine operation data monitoring system has achieved multi-dimensional and full-scenario coverage, and the collection of multi-dimensional data such as vibration and speed has provided a rich foundation for life prediction. However, how to reduce noise interference and obtain better prediction results remains the core challenge of wind turbine life prediction. Since the measured data of wind turbines often contain a large amount of noise and anomalies, if these interferences are not effectively processed, they will seriously affect the prediction effect. In addition, the occurrence of wind power failures is usually accompanied by a variety of complex mechanisms, involving the mutual coupling effects between multiple components, and a single analysis method is difficult to fully cover these factors. Summary of the Invention
[0003] In order to overcome the shortcomings of the prior art, the present invention aims to provide a wind turbine life prediction method based on empirical mode decomposition (EMD) and hybrid neural networks, thereby improving the accuracy of wind turbine life prediction.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A wind turbine life prediction method based on empirical mode decomposition and hybrid neural network is proposed. First, empirical mode decomposition is performed on the effective value of the original vibration signal of the wind turbine to obtain several intrinsic mode functions (IMFs). Then, a hybrid neural network is used to predict each IMF component. Finally, the prediction results of each IMF are weighted and fused to obtain the overall prediction result of the original vibration signal.
[0005] The wind turbine lifetime prediction method based on empirical mode decomposition and hybrid neural networks includes the following steps: The first step is to preprocess the raw vibration signal dataset collected by the wind turbine's sensors. First, the data is cleaned to remove outliers, missing values, and invalid redundant data. Then, noise interference is removed through filtering algorithms to extract the effective values of the raw vibration signals. Finally, the operating condition is normalized to convert the actual speed into a unified reference speed, thus achieving standardized alignment of data from multiple operating conditions. The second step employs a degenerate trajectory prediction method based on adaptive noise complete set empirical mode decomposition: white noise is added to the original vibration signal, a new signal sequence is constructed, and the amplitude and frequency distribution of the white noise are dynamically adjusted through an adaptive algorithm. Then, the signal after adding white noise is analyzed... Multiple CEEMDAN decompositions are performed to obtain multiple sets of intrinsic mode functions; in, It is the first The j-th IMF component extracted in the nth iteration Represents the residual; The third step employs a CNN-LSTM hybrid neural network model based on an attention mechanism. This model is used to predict each intrinsic mode component sequence obtained from the decomposition in the second step, thereby generating the corresponding prediction sequence. The CNN-LSTM hybrid neural network model consists of a two-layer CNN, a two-layer LSTM, an attention layer, and an output layer. Multi-scale convolutional feature extraction constructs time-sensitive vectors by applying a feature fusion mechanism after two layers of convolution and pooling operations. The input data dimension is 50×25, representing 50 historical time steps. Each time step contains a 25-dimensional feature vector, which is adapted to the spatiotemporal characteristics of multi-dimensional time series. The first convolutional layer uses 32 1×3 convolutional kernels to mine local patterns across feature dimensions. This is the output feature map of the first convolutional layer; These are the kernel weight parameters for the first convolutional layer; : Bias parameters of the first convolutional layer; Then, the temporal dimension is compressed to 25×32 using 1×2 max pooling: The first convolutional layer outputs a feature map after pooling; pool() represents the pooling operation function.
[0006] The second convolutional layer uses 16 1×3 convolutional kernels for deep feature abstraction, with an output dimension of 25×16. This is the output feature map of the second convolutional layer; These are the kernel weight parameters for the second convolutional layer; : Bias parameters of the second convolutional layer; Then, pooling is performed, and finally, the pooled feature map is flattened and mapped to a temporally sensitive vector through a fully connected layer: In the formula, F={ The output time-sensitive vector will be used as the input feature for subsequent models. Let f(c) be the weight parameter matrix of the fully connected layer; flatten() represents the feature flattening operation function; b(fc) is the bias parameter vector of the fully connected layer. The second convolutional layer outputs the pooled feature map; Subsequently, a temporal state transfer module was established, utilizing the F vector to achieve dimensionality adaptation between CNN and LSTM, through cell states. Achieving long-term memory updates: = In the formula ( ) represents the forward computation function of the LSTM unit; This represents the eigenvector at time step t; Indicates the previous moment The LSTM state; And it dynamically adjusts the intensity of information forgetting and memory through a gating mechanism; For multi-attention computation, the LSTM hidden state is split into h parallel subspaces, and the i-th attention head... The operation is as follows: In the formula This is the query, key, and value matrix after linear projection; This is the scaling factor used to scale the dot product result.
[0007] For sub-attention focusing, firstly, for each time step t, calculate its correlation score with other time steps: In the formula For time step and The correlation score For time step The key vector, It is the query vector at time step t; These scores are normalized using the softmax function at time steps. attention weights : Using these weights, a weighted summation vector is obtained to obtain a new time step representation; This represents the weight parameter matrix of the multi-head attention output layer; This represents the bias parameter vector of the multi-head attention output layer; This is the output of the i-th attention head. Indicates the prediction step size; It represents the total number of attention heads.
[0008] The fourth step is to synthesize the prediction signal. All the prediction sequences obtained in the third step are linearly superimposed and synthesized to generate the final synthesized signal, which is the prediction output. Finally, the prediction result is obtained; that is, the fully connected layer maps the fused features to the prediction space. in This represents the signal predicted after n iterations. It is the predicted sequence of the j-th IMF component after n iterations; This represents the residual components in the corresponding iteration step.
[0009] The fifth step is to construct and output a life prediction model. By analyzing historical fault data of wind turbine units, combined with the equipment's factory technical parameters and industry standards, the degradation failure threshold corresponding to the vibration signal is determined. Predicted sequence based on synthesized vibration signal The remaining useful life (RUL) is calculated using the trend extrapolation method; the specific steps are as follows: By performing trend fitting on the predicted sequence, the degradation trend equation is obtained: Solve the equation The predicted failure time is obtained. , Calculate remaining lifetime: in This is the current forecast start time.
[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively addresses the shortcomings of traditional life prediction methods in wind turbine life prediction, such as insufficient noise resistance and poor adaptability to complex operating conditions, by combining an improved empirical mode decomposition (CEEMDAN) and a hybrid neural network (CNN-LSTM + Attention Mechanism). This method not only significantly improves prediction accuracy but also enhances the model's robustness and interpretability, providing strong support for intelligent operation and maintenance in the wind power industry. Attached Figure Description
[0011] Figure 1 This is a flowchart of an embodiment of the present invention.
[0012] Figure 2 This is a schematic diagram of the hybrid neural network structure according to an embodiment of the present invention.
[0013] Figure 3 These are the prediction results for some IMF components in an embodiment of the present invention.
[0014] Figure 4 This is the lifetime prediction result of an embodiment of the present invention. Detailed Implementation
[0015] The present invention will be further described below with reference to embodiments and accompanying drawings.
[0016] Reference Figure 1 A method for predicting the lifespan of wind turbine generators based on empirical mode decomposition and hybrid neural networks includes the following steps: The first step is to preprocess the raw vibration signal dataset collected by the wind turbine's sensors. First, data cleaning is performed to remove outliers, missing values, and invalid redundant data. Then, filtering algorithms are used to remove noise interference and improve signal quality. Based on this, key features such as the effective values of the vibration signals are extracted, and operating condition normalization is performed to convert the actual rotational speed into a unified reference speed. This standardization and alignment of data from multiple operating conditions improves the overall reliability and usability of the dataset. The second step, in order to effectively describe complex degenerate trajectories that include overall trends and local fluctuations, while reducing noise interference, is to propose a degenerate trajectory prediction method based on adaptive noise complete set empirical mode decomposition.
[0017] A certain amount of white noise is added to the original vibration signal to construct a new signal sequence. An adaptive algorithm dynamically adjusts the amplitude and frequency distribution of the white noise to ensure uniform noise distribution and stability during the decomposition process. Then, the signal after adding white noise is analyzed... Multiple CEEMDAN decompositions are performed to obtain multiple sets of intrinsic mode functions.
[0018] in, It is the first The j-th IMF component extracted in the nth iteration Represents the residual; The third step employs a CNN-LSTM hybrid neural network model based on an attention mechanism. This model combines the powerful abstract information extraction capabilities of convolutional neural networks with the long-term memory function of long short-term memory neural networks, enabling it to effectively handle complex time-series data. Furthermore, by introducing a self-attention mechanism, the model can more accurately focus on key features in the time series, thereby significantly improving prediction accuracy. In this invention, this hybrid neural network model is used to predict each intrinsic mode component sequence obtained from the decomposition in the second step, thereby generating the corresponding prediction sequence. This neural network model includes a two-layer CNN, a two-layer LSTM, an attention layer, and an output layer. The network structure diagram is shown below. Figure 2 As shown; Multi-scale convolutional feature extraction constructs time-sensitive vectors by applying a feature fusion mechanism after two layers of convolution and pooling operations. The input data dimension is 50×25, representing 50 historical time steps. Each time step contains a 25-dimensional feature vector, which is adapted to the spatiotemporal characteristics of multi-dimensional time series. The first convolutional layer uses 32 1×3 convolutional kernels to mine local patterns across feature dimensions. This is the output feature map of the first convolutional layer; These are the kernel weight parameters for the first convolutional layer; : Bias parameters of the first convolutional layer; Then, the temporal dimension is compressed to 25×32 using 1×2 max pooling: The first convolutional layer outputs a feature map after pooling; pool() represents the pooling operation function.
[0019] The second convolutional layer uses 16 1×3 convolutional kernels for deep feature abstraction, with an output dimension of 25×16. This is the output feature map of the second convolutional layer; These are the kernel weight parameters for the second convolutional layer; : Bias parameters of the second convolutional layer; Then, pooling is performed, and finally, the pooled feature map is flattened and mapped to a temporally sensitive vector through a fully connected layer: In the formula, F={ The output time-sensitive vector will be used as the input feature for subsequent models. Let f(c) be the weight parameter matrix of the fully connected layer; flatten() represents the feature flattening operation function; b(fc) is the bias parameter vector of the fully connected layer. The second convolutional layer outputs the pooled feature map; Subsequently, a temporal state transfer module was established, utilizing the F vector to achieve dimensionality adaptation between CNN and LSTM, through cell states. Achieving long-term memory updates: = In the formula ( ) represents the forward computation function of the LSTM cell. This represents the eigenvector at time step t; Indicates the previous moment The LSTM state; It also dynamically adjusts the intensity of information forgetting and memory through a gating mechanism.
[0020] In this architecture, a sub-attention mechanism is applied to the output of the LSTM to better capture important information in the time series. For multi-attention computation, the LSTM hidden state is split into h parallel subspaces, and the operation of the i-th head is: In the formula This is the query, key, and value matrix after linear projection; This is the scaling factor used to scale the dot product result.
[0021] For sub-attention focusing, firstly, for each time step t, calculate its correlation score with other time steps: In the formula For time step and The correlation score For time step The key vector, It is the query vector at time step t; These scores are normalized using the softmax function at time steps. attention weights : Using these weights, a weighted summation vector is obtained to obtain a new time step representation.
[0022] This represents the weight parameter matrix of the multi-head attention output layer; This represents the bias parameter vector of the multi-head attention output layer; This is the output of the i-th attention head. Indicates the prediction step size; It represents the total number of attention heads.
[0023] The fourth step is to synthesize the prediction signal. All the prediction sequences obtained in the third step are linearly superimposed and synthesized to generate the final synthesized signal, which is the prediction output. Finally, the prediction result is obtained; that is, the fully connected layer maps the fused features to the prediction space. in This represents the signal predicted after n iterations. It is the predicted sequence of the j-th IMF component after n iterations; This represents the residual components in the corresponding iteration step.
[0024] The fifth step is to construct and output a life prediction model. By analyzing historical fault data of wind turbine units, combined with the equipment's factory technical parameters and industry standards, the degradation failure threshold corresponding to the vibration signal is determined. Predicted sequence based on synthesized vibration signal The remaining useful life (RUL) is calculated using the trend extrapolation method. The specific steps are as follows: By performing trend fitting on the predicted sequence, the degradation trend equation is obtained: Solve the equation The predicted failure time is obtained. , Calculate remaining lifetime: in This is the current forecast start time.
[0025] Reference Figure 3 , Figure 3The following are prediction results for some IMF components in embodiments of the present invention. It can be seen that the model achieves high-precision prediction for IMF components with different frequency characteristics. The predicted curves on the training and test sets generally match the actual curves well, effectively capturing both local fluctuations and overall trends of the signal, thus verifying the model's ability to predict the time series of non-stationary vibration signal components. (Refer to...) Figure 4 , Figure 4 The lifetime prediction results of the embodiments of the present invention show that the RMS degradation trajectory predicted by the method of the present invention is basically consistent with the actual degradation trend. The predicted lifetime endpoint has a small deviation from the actual failure node, which can effectively realize the prediction of the remaining lifetime of wind turbine units and provide a reliable basis for equipment maintenance decisions.
Claims
1. A wind turbine lifetime prediction method based on empirical mode decomposition and hybrid neural network, characterized in that: First, empirical mode decomposition is performed on the effective value of the original vibration signal of the wind turbine to obtain several intrinsic mode functions (IMFs). Then, a hybrid neural network is used to predict each IMF component. Finally, the prediction results of each IMF are weighted and fused to obtain the overall prediction result of the original vibration signal.
2. The wind turbine lifetime prediction method of claim 1, wherein, Includes the following steps: The first step is to perform data preprocessing on the raw vibration signal dataset collected by the wind turbine's sensors; The second step involves processing the original vibration signal using a degenerate trajectory prediction method based on adaptive noise complete set empirical mode decomposition. The third step involves using a CNN-LSTM hybrid neural network model based on an attention mechanism to predict each intrinsic mode component sequence obtained from the decomposition in the second step, thereby generating the corresponding prediction sequence. The fourth step involves linearly superimposing all the predicted sequences obtained in the third step to synthesize the final synthesized signal, thus generating the prediction output. Finally, the prediction result is obtained; that is, the fully connected layer maps the fused features to the prediction space. wherein represents the predicted signal after n iterations; is the predicted sequence of the jth IMF component after n iterations; is the residual component in the corresponding iteration step; The fifth step is to construct and output a life prediction model. By analyzing historical fault data of wind turbine units and combining the equipment's factory technical parameters with industry standards, the degradation failure threshold corresponding to the vibration signal is determined. By fitting the degradation trend equation of the predicted sequence through trend extrapolation, the failure time is solved and the remaining life of the wind turbine is calculated.
3. The wind turbine life prediction method according to claim 2, characterized in that, The first step is as follows: First, data cleaning is performed to remove outliers, missing values, and invalid redundant data; then, noise interference is removed through filtering algorithms, the effective values of the original vibration signals are extracted, and the working condition normalization process is completed to convert the actual speed into a unified reference speed, thereby achieving standardized alignment of multi-working condition data.
4. The wind turbine lifetime prediction method of claim 2, wherein, The second step specifically involves: adding white noise to the original vibration signal, constructing a new signal sequence, dynamically adjusting the amplitude and frequency distribution of the white noise using an adaptive algorithm, and then processing the signal after adding white noise. Multiple CEEMDAN decompositions are performed to obtain multiple sets of intrinsic mode functions; wherein, is the jth IMF component extracted in the nth iteration, is the jth IMF component extracted in the nth iteration, denotes the residual.
5. The wind turbine lifetime prediction method of claim 2, wherein, The third step specifically involves: the CNN-LSTM hybrid neural network model comprising a two-layer CNN, a two-layer LSTM, an attention layer, and an output layer; Multi-scale convolutional feature extraction constructs time-sensitive vectors by applying a feature fusion mechanism after two layers of convolution and pooling operations. The input data dimension is 50×25, representing 50 historical time steps. Each time step contains a 25-dimensional feature vector, which is adapted to the spatiotemporal characteristics of multi-dimensional time series. The first convolutional layer uses 32 1×3 convolutional kernels to mine local patterns across feature dimensions. = This is the output feature map of the first convolutional layer; These are the kernel weight parameters for the first convolutional layer; : Bias parameters of the first convolutional layer; Then, the temporal dimension is compressed to 25×32 using 1×2 max pooling: The first convolutional layer outputs the pooled feature map, and pool() represents a pooling operation function. The second convolutional layer uses 16 1×3 convolutional kernels for deep feature abstraction, with an output dimension of 25×16. is an output feature map of the second convolutional layer; is a convolution kernel weight parameter of the second convolutional layer; is a bias parameter of the second convolutional layer; Then, pooling is performed, and finally, the pooled feature map is flattened and mapped to a temporally sensitive vector through a fully connected layer: In the formula, F={ The output time-sensitive vector will be used as the input feature for subsequent models. is the weight parameter matrix of the fully connected layer; flatten() represents the feature flattening operation function; b(fc) is the bias parameter vector of the fully connected layer; The second convolutional layer outputs the pooled feature map; Then the timing state transfer module is established, the dimension adaptation of CNN and LSTM is realized by using F vector, the long-term memory update is realized by cell state implementation = In the formula ( ) is a forward computation function of the LSTM unit; represents a feature vector at time step t; represents an LSTM state at the previous moment . The intensity of information forgetting and memory is dynamically adjusted through a gating mechanism; For multi-attention computation, the LSTM hidden state is split into h parallel subspaces, the i-th attention head is computed as: wherein is the linear projected query, key, value matrix; is a scaling factor used to scale the dot product result; For sub-attention focusing, firstly, for each time step t, calculate its correlation score with other time steps: In the formula For time steps and The correlation score For time steps The key vector; It is the query vector at time step t; These scores are normalized using the softmax function to obtain the time steps. Attention weights : Using these weights, a weighted summation vector is obtained to obtain a new time step representation; a weight parameter matrix representing the multi-headed attention output layer; a bias parameter vector representing the multi-headed attention output layer; is an output result for the i-th attention head; represents a prediction step; is a total number of attention heads.
6. The wind turbine lifetime prediction method of claim 2, wherein, The fifth step specifically involves: performing trend fitting on the predicted sequence to obtain the degradation trend equation. solving the equation to obtain a predicted time to failure , Calculate remaining lifetime: wherein is the current prediction start time.